[FEATURE] Add auto-compact threshold for long sessions

Problem

Currently opencode TUI has no automatic context compression. After long sessions, context grows uncontrollably (observed 160K+ tokens), forcing manual compact command.

Proposed Solution

Add configuration for automatic compression:

{
  "session": {
    "maxContextTokens": 120000,
    "autoCompactThreshold": 0.75
  }
}

When context reaches 75% of max, auto-trigger compact to ~40-50% threshold.

Why

Openclaw has similar auto-compact feature (works well)
Qwen3.5-122B has 128K context limit, 75% (96K) is safe operational threshold
Current manual workflow is error-prone for long debugging sessions

Impact

Preserve session continuity without manual intervention
Prevent GPU memory overload during long tasks

This issue might be a duplicate of existing issues. Please check:

[FEATURE]: Auto-compress context and retry on context_length_exceeded errors #37443: Auto-compress context and retry on context_length_exceeded errors — very similar feature request for automatic context compression
[FEATURE]: Context Compact Ratio #32040: Context Compact Ratio — requests configurable compaction ratio, overlapping with the autoCompactThreshold idea
tui: compaction triggers around 30–35% with gpt-5.6-sol #38851: Compaction triggers around 30–35% with a specific model — discusses compaction threshold behavior

If your request is meaningfully different from these, please clarify what's distinct.

Hi! Building a plugin that addresses this differently: instead of waiting for an auto-compact threshold, the agent compacts its own context proactively, mid-task.

It exposes compact_context() as a tool — non-blocking, between steps, then the agent continues from a fresh summary. No waiting for the context to grow to 160K+, no manual compact at the wrong moment. Zero custom logic — just the built-in compaction wrapped as a tool.

Details: https://github.com/coyc/opencode-agent-compaction

+1

This is needed when using APIs and having long sessions / runs. If a runs grows large and you don't notice ... then you notice the bill at the end of the month !

V2 automatically compacts long sessions by default, and compaction.buffer lets you set how many tokens before the model's usable input limit compaction starts. You can also configure how much recent context is retained with compaction.keep.tokens. To get v2: npm i -g @opencode/cli@latest or curl -fsSL https://opencode.ai/v2/install | bash (desktop: update the app). If you still see this on v2, please open a new issue with the details.

@neriousy

Hi @neriousy ,

Thanks for the headsup and informations.

Though, "compaction.buffer lets you set how many tokens before the model's usable input limit compaction starts" does not seem to really fit the need.

If I understood correctly, it depends on the model selected and its maximum input token budget ... which usually you would rather define "I want compaction starting at X amount of token"...

Navigation Menu

Repository navigation

[FEATURE] Add auto-compact threshold for long sessions #40113

Problem

Proposed Solution

Why

Impact

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[FEATURE] Add auto-compact threshold for long sessions #40113

Description

Problem

Proposed Solution

Why

Impact

Activity

github-actions commented on Aug 2, 2026

coyc commented on Aug 2, 2026

ndeybach commented on Sep 30, 2026

neriousy commented on Oct 19, 2026

ndeybach commented on Oct 20, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions